Virtualization - Apr 2023 - [RFC PATCH net 2/3] virtio-net: allow usage of vrings smaller than MAX_SKB_FRAGS + 2

On Sun, Apr 30, 2023 at 04:15:17PM +0300, Alvaro Karsz
wrote:
> At the moment, if a network device uses vrings with less than
> MAX_SKB_FRAGS + 2 entries, the device won't be functional.
> 
> The following condition vq->num_free >= 2 + MAX_SKB_FRAGS will always
> evaluate to false, leading to TX timeouts.
> 
> This patch introduces a new variable, single_pkt_max_descs, that holds
> the max number of descriptors we may need to handle a single packet.
> 
> This patch also detects the small vring during probe, blocks some
> features that can't be used with small vrings, and fails probe,
> leading to a reset and features re-negotiation.
> 
> Features that can't be used with small vrings:
> GRO features (VIRTIO_NET_F_GUEST_*):
> When we use small vrings, we may not have enough entries in the ring to
> chain page size buffers and form a 64K buffer.
> So we may need to allocate 64k of continuous memory, which may be too
> much when the system is stressed.
> 
> This patch also fixes the MTU size in small vring cases to be up to the
> default one, 1500B.
and then it should clear VIRTIO_NET_F_MTU?
> Signed-off-by: Alvaro Karsz <alvaro.karsz at solid-run.com>


> ---
>  drivers/net/virtio_net.c | 149 +++++++++++++++++++++++++++++++++++++--
>  1 file changed, 144 insertions(+), 5 deletions(-)
> 
> diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
> index 8d8038538fc..b4441d63890 100644
> --- a/drivers/net/virtio_net.c
> +++ b/drivers/net/virtio_net.c
> @@ -103,6 +103,8 @@ struct virtnet_rq_stats {
>  #define VIRTNET_SQ_STAT(m)	offsetof(struct virtnet_sq_stats, m)
>  #define VIRTNET_RQ_STAT(m)	offsetof(struct virtnet_rq_stats, m)
>  
> +#define IS_SMALL_VRING(size)	((size) < MAX_SKB_FRAGS + 2)
> +
>  static const struct virtnet_stat_desc virtnet_sq_stats_desc[] = {
>  	{ "packets",		VIRTNET_SQ_STAT(packets) },
>  	{ "bytes",		VIRTNET_SQ_STAT(bytes) },
> @@ -268,6 +270,12 @@ struct virtnet_info {
>  	/* Does the affinity hint is set for virtqueues? */
>  	bool affinity_hint_set;
>  
> +	/* How many ring descriptors we may need to transmit a single packet */
> +	u16 single_pkt_max_descs;
> +
> +	/* Do we have virtqueues with small vrings? */
> +	bool svring;
> +
>  	/* CPU hotplug instances for online & dead */
>  	struct hlist_node node;
>  	struct hlist_node node_dead;
worth checking that all these layout changes don't push useful things to
a different cache line. can you add that analysis?

I see confusiong here wrt whether some rings are "small"? all of them?
some rx rings? some tx rings? names should make it clear.
also do we really need bool svring? can't we just check single_pkt_max_descs
all the time?
> @@ -455,6 +463,7 @@ static struct sk_buff *page_to_skb(struct virtnet_info
*vi,
>  	unsigned int copy, hdr_len, hdr_padded_len;
>  	struct page *page_to_free = NULL;
>  	int tailroom, shinfo_size;
> +	u16 max_frags = MAX_SKB_FRAGS;
>  	char *p, *hdr_p, *buf;
>  
>  	p = page_address(page) + offset;
> @@ -520,7 +529,10 @@ static struct sk_buff *page_to_skb(struct virtnet_info
*vi,
>  	 * tries to receive more than is possible. This is usually
>  	 * the case of a broken device.
>  	 */
> -	if (unlikely(len > MAX_SKB_FRAGS * PAGE_SIZE)) {
> +	if (unlikely(vi->svring))
> +		max_frags = 1;
> +
> +	if (unlikely(len > max_frags * PAGE_SIZE)) {
>  		net_dbg_ratelimited("%s: too much data\n",
skb->dev->name);
>  		dev_kfree_skb(skb);
>  		return NULL;
> @@ -612,7 +624,7 @@ static void check_sq_full_and_disable(struct
virtnet_info *vi,
>  	 * Since most packets only take 1 or 2 ring slots, stopping the queue
>  	 * early means 16 slots are typically wasted.
>  	 */
> -	if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
> +	if (sq->vq->num_free < vi->single_pkt_max_descs) {
>  		netif_stop_subqueue(dev, qnum);
>  		if (use_napi) {
>  			if (unlikely(!virtqueue_enable_cb_delayed(sq->vq)))
> @@ -620,7 +632,7 @@ static void check_sq_full_and_disable(struct
virtnet_info *vi,
>  		} else if (unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
>  			/* More just got used, free them then recheck. */
>  			free_old_xmit_skbs(sq, false);
> -			if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
> +			if (sq->vq->num_free >= vi->single_pkt_max_descs) {
>  				netif_start_subqueue(dev, qnum);
>  				virtqueue_disable_cb(sq->vq);
>  			}
> @@ -1108,6 +1120,10 @@ static int virtnet_build_xdp_buff_mrg(struct
net_device *dev,
>  		return 0;
>  
>  	if (*num_buf > 1) {
> +		/* Small vring - can't be more than 1 buffer */
> +		if (unlikely(vi->svring))
> +			return -EINVAL;
> +
>  		/* If we want to build multi-buffer xdp, we need
>  		 * to specify that the flags of xdp_buff have the
>  		 * XDP_FLAGS_HAS_FRAG bit.
> @@ -1828,7 +1844,7 @@ static void virtnet_poll_cleantx(struct receive_queue
*rq)
>  			free_old_xmit_skbs(sq, true);
>  		} while (unlikely(!virtqueue_enable_cb_delayed(sq->vq)));
>  
> -		if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
> +		if (sq->vq->num_free >= vi->single_pkt_max_descs)
>  			netif_tx_wake_queue(txq);
>  
>  		__netif_tx_unlock(txq);
> @@ -1919,7 +1935,7 @@ static int virtnet_poll_tx(struct napi_struct *napi,
int budget)
>  	virtqueue_disable_cb(sq->vq);
>  	free_old_xmit_skbs(sq, true);
>  
> -	if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
> +	if (sq->vq->num_free >= vi->single_pkt_max_descs)
>  		netif_tx_wake_queue(txq);
>  
>  	opaque = virtqueue_enable_cb_prepare(sq->vq);
> @@ -3862,6 +3878,15 @@ static bool virtnet_check_guest_gso(const struct
virtnet_info *vi)
>  		virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_USO6));
>  }
>  
> +static bool virtnet_check_host_gso(const struct virtnet_info *vi)
> +{
> +	return virtio_has_feature(vi->vdev, VIRTIO_NET_F_HOST_TSO4) ||
> +		virtio_has_feature(vi->vdev, VIRTIO_NET_F_HOST_TSO6) ||
> +		virtio_has_feature(vi->vdev, VIRTIO_NET_F_HOST_ECN) ||
> +		virtio_has_feature(vi->vdev, VIRTIO_NET_F_HOST_UFO) ||
> +		virtio_has_feature(vi->vdev, VIRTIO_NET_F_HOST_USO);
> +}
> +
>  static void virtnet_set_big_packets(struct virtnet_info *vi, const int
mtu)
>  {
>  	bool guest_gso = virtnet_check_guest_gso(vi);
> @@ -3876,6 +3901,112 @@ static void virtnet_set_big_packets(struct
virtnet_info *vi, const int mtu)
>  	}
>  }
>  
> +static u16 virtnet_calc_max_descs(struct virtnet_info *vi)
> +{
> +	if (vi->svring) {
> +		if (virtnet_check_host_gso(vi))
> +			return 4; /* 1 fragment + linear part + virtio header + GSO header */
> +		else
> +			return 3; /* 1 fragment + linear part + virtio header */
> +	} else {
> +		return MAX_SKB_FRAGS + 2;
> +	}
> +}
> +
> +static bool virtnet_uses_svring(struct virtnet_info *vi)
> +{
> +	u32 i;
> +
> +	/* If a transmit/receive virtqueue is small,
> +	 * we cannot handle fragmented packets.
> +	 */
> +	for (i = 0; i < vi->max_queue_pairs; i++) {
> +		if (IS_SMALL_VRING(virtqueue_get_vring_size(vi->sq[i].vq)) ||
> +		    IS_SMALL_VRING(virtqueue_get_vring_size(vi->rq[i].vq)))
> +			return true;
> +	}
> +
> +	return false;
> +}
I see even if only some rings are too small we force everything to use
small ones. Wouldn't it be better to just disable small ones in this
case? That would not need a reset.

> +
> +/* Function returns the number of features it blocked */
We don't need the # though. Make it bool?
> +static int virtnet_block_svring_unsupported(struct virtio_device *vdev)
> +{
> +	int cnt = 0;
> +	/* Block Virtio guest GRO features.
> +	 * Asking Linux to allocate 64k of continuous memory is too much,
> +	 * specially when the system is stressed.
> +	 *
> +	 * If VIRTIO_NET_F_MRG_RXBUF is negotiated we can allcoate smaller
> +	 * buffers, but since the ring is small, the buffers can be quite big.
> +	 *
> +	 */
> +	if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4)) {
> +		virtio_block_feature(vdev, VIRTIO_NET_F_GUEST_TSO4);
> +		cnt++;
> +	}
> +	if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6)) {
> +		virtio_block_feature(vdev, VIRTIO_NET_F_GUEST_TSO6);
> +		cnt++;
> +	}
> +	if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN)) {
> +		virtio_block_feature(vdev, VIRTIO_NET_F_GUEST_ECN);
> +		cnt++;
> +	}
> +	if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_UFO)) {
> +		virtio_block_feature(vdev, VIRTIO_NET_F_GUEST_UFO);
> +		cnt++;
> +	}
> +	if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_USO4)) {
> +		virtio_block_feature(vdev, VIRTIO_NET_F_GUEST_USO4);
> +		cnt++;
> +	}
> +	if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_USO6)) {
> +		virtio_block_feature(vdev, VIRTIO_NET_F_GUEST_USO6);
> +		cnt++;
> +	}
> +
> +	return cnt;
> +}
> +
> +static int virtnet_fixup_svring(struct virtnet_info *vi)
> +{
> +	int i;
> +	/* Do we use small vrings?
> +	 * If not, nothing we need to do.
> +	 */
> +	vi->svring = virtnet_uses_svring(vi);
> +	if (!vi->svring)
> +		return 0;
> +
> +	/* Some features can't be used with small vrings.
> +	 * Block those and return an error.
> +	 * This will trigger a reprobe without the blocked
> +	 * features.
> +	 */
> +	if (virtnet_block_svring_unsupported(vi->vdev))
> +		return -EOPNOTSUPP;
> +
> +	/* Disable NETIF_F_SG */
> +	vi->dev->hw_features &= ~NETIF_F_SG;
> +
> +	/* Don't use MTU bigger than default */
> +	if (vi->dev->max_mtu > ETH_DATA_LEN)
> +		vi->dev->max_mtu = ETH_DATA_LEN;
> +	if (vi->dev->mtu > ETH_DATA_LEN)
> +		vi->dev->mtu = ETH_DATA_LEN;
> +
> +	/* Don't use big packets */
> +	vi->big_packets = false;
> +	vi->big_packets_num_skbfrags = 1;
> +
> +	/* Fix min_buf_len for receive virtqueues */
> +	for (i = 0; i < vi->max_queue_pairs; i++)
> +		vi->rq[i].min_buf_len = mergeable_min_buf_len(vi, vi->rq[i].vq);
> +
> +	return 0;
> +}
> +
>  static int virtnet_probe(struct virtio_device *vdev)
>  {
>  	int i, err = -ENOMEM;
> @@ -4061,6 +4192,14 @@ static int virtnet_probe(struct virtio_device *vdev)
>  	if (err)
>  		goto free;
>  
> +	/* Do required fixups in case we are using small vrings */
> +	err = virtnet_fixup_svring(vi);
> +	if (err)
> +		goto free_vqs;
> +
> +	/* Calculate the max. number of descriptors we may need to transmit a
single packet */
> +	vi->single_pkt_max_descs = virtnet_calc_max_descs(vi);
> +
>  #ifdef CONFIG_SYSFS
>  	if (vi->mergeable_rx_bufs)
>  		dev->sysfs_rx_queue_group = &virtio_net_mrg_rx_group;
> -- 
> 2.34.1

> > At the moment, if a network device uses vrings with less than
> > MAX_SKB_FRAGS + 2 entries, the device won't be functional.
> >
> > The following condition vq->num_free >= 2 + MAX_SKB_FRAGS will
always
> > evaluate to false, leading to TX timeouts.
> >
> > This patch introduces a new variable, single_pkt_max_descs, that holds
> > the max number of descriptors we may need to handle a single packet.
> >
> > This patch also detects the small vring during probe, blocks some
> > features that can't be used with small vrings, and fails probe,
> > leading to a reset and features re-negotiation.
> >
> > Features that can't be used with small vrings:
> > GRO features (VIRTIO_NET_F_GUEST_*):
> > When we use small vrings, we may not have enough entries in the ring
to
> > chain page size buffers and form a 64K buffer.
> > So we may need to allocate 64k of continuous memory, which may be too
> > much when the system is stressed.
> >
> > This patch also fixes the MTU size in small vring cases to be up to
the
> > default one, 1500B.
> 
> and then it should clear VIRTIO_NET_F_MTU?
> 
Following [1], I was thinking to accept the feature and a let the device figure
out that it can't transmit a big packet, since the RX buffers are not big
enough (without VIRTIO_NET_F_MRG_RXBUF).
But, I think that we may need to block the MTU feature after all.
Quoting the spec:

A driver SHOULD negotiate VIRTIO_NET_F_MTU if the device offers it.
If the driver negotiates VIRTIO_NET_F_MTU, it MUST supply enough receive buffers
to receive at least one receive packet of size mtu (plus low level ethernet
header length) with gso_type NONE or ECN.

So, if VIRTIO_NET_F_MTU is negotiated, we MUST supply enough receive buffers.
So I think that blocking VIRTIO_NET_F_MTU  should be the way to go, If mtu >
1500.

[1] https://lore.kernel.org/lkml/20230417031052-mutt-send-email-mst at
kernel.org/
> > +     /* How many ring descriptors we may need to transmit a single
packet */
> > +     u16 single_pkt_max_descs;
> > +
> > +     /* Do we have virtqueues with small vrings? */
> > +     bool svring;
> > +
> >       /* CPU hotplug instances for online & dead */
> >       struct hlist_node node;
> >       struct hlist_node node_dead;
> 
> worth checking that all these layout changes don't push useful things
to
> a different cache line. can you add that analysis?
> 
Good point.
I think that we can just move these to the bottom of the struct.
> 
> I see confusiong here wrt whether some rings are "small"? all of
them?
> some rx rings? some tx rings? names should make it clear.
The small vring is a device attribute, not a vq attribute. It blocks features,
which affects the entire device.
Maybe we can call it "small vring mode".
> also do we really need bool svring? can't we just check
single_pkt_max_descs
> all the time?
> 
We can work without the bool, we could always check if single_pkt_max_descs !=
MAX_SKB_FRAGS + 2.
It doesn't really matter to me, I was thinking it may be more readable this
way.
> > +static bool virtnet_uses_svring(struct virtnet_info *vi)
> > +{
> > +     u32 i;
> > +
> > +     /* If a transmit/receive virtqueue is small,
> > +      * we cannot handle fragmented packets.
> > +      */
> > +     for (i = 0; i < vi->max_queue_pairs; i++) {
> > +             if
(IS_SMALL_VRING(virtqueue_get_vring_size(vi->sq[i].vq)) ||
> > +                
IS_SMALL_VRING(virtqueue_get_vring_size(vi->rq[i].vq)))
> > +                     return true;
> > +     }
> > +
> > +     return false;
> > +}
> 
> I see even if only some rings are too small we force everything to use
> small ones. Wouldn't it be better to just disable small ones in this
> case? That would not need a reset.
> 
I'm not sure. It may complicate things.

What if all TX vqs are small?
What if all RX vqs are small?
What if we end up with an unbalanced number of TX vqs and RX vqs? is this
allowed by the spec?
What if we end up disabling the RX default vq (receiveq1)?

I guess we could do it, after checking some conditions.
Maybe we can do it in a follow up patch?
Do you think it's important for it to be included since day 1?

I think that the question is: what's more important, to use all the vqs
while blocking some features, or to use part of the vqs without blocking
features?
> > +
> > +/* Function returns the number of features it blocked */
> 
> We don't need the # though. Make it bool?
> 
Sure.